Research / ear-eeg-emg-separation
Separating and jointly modelling EEG and EMG under reduced, ear-based sensor configurations
Measures how decoding performance changes when EEG- and EMG-related information are separated and then modelled together, as the sensor array is reduced toward an ear-based configuration.
How does separating and jointly modelling EEG- and EMG-related information affect decoding performance under reduced, ear-based wearable sensor configurations?
A wearable BCI cannot carry a full scalp montage, so the practical question is not whether decoding works with 64 electrodes but how fast it degrades on the way down to four or eight. Standard pipelines remove muscle activity as an artefact. On a device sitting beside the jaw, that activity is both unavoidable and possibly informative, which makes "remove it" an assumption worth testing rather than a default worth inheriting.
Around-ear and in-ear EEG has an established literature covering electrode design, achievable signal quality and the tasks that survive reduced coverage. Separately, there is work on silent-speech and articulation-related decoding where facial EMG carries substantial information. The gap this study addresses sits between them: reduced ear-proximal configurations evaluated with EEG and EMG contributions deliberately separated and then recombined, rather than one treated as noise in the other.
- Data
- Public EEG/EMG datasets. Specific datasets to be fixed at pre-registration.
- Channel configurations
- Full montage baseline, progressively reduced subsets, ear-proximal subset.
- Preprocessing
- Band-pass filtering, re-referencing, artefact handling with and without EMG suppression.
- Separation
- Source-separation and band-limited decomposition to isolate EEG- and EMG-dominant components.
- Models
- Classical baseline, convolutional baseline, and a fusion model over the separated streams.
- Evaluation
- Subject-dependent and subject-independent splits, reported separately.
- Ablations
- EEG-only, EMG-only, naive-combined, separated-and-fused, across each channel configuration.
- Step 01Fix task, datasets and metrics before any modelling.
- Step 02Establish a full-montage baseline for each model family.
- Step 03Reduce the channel set in stages, re-evaluating at each step.
- Step 04Repeat with EMG suppressed, EMG retained, and EMG modelled as a separate stream.
- Step 05Report the performance-versus-sensor-count curve for each condition.
No results.
This study has not produced data. Figures, tables and statistics will appear here once analysis is complete and only then.
Cerebral does not publish placeholder numbers. An empty results section is the honest state of a study in design.
- L01Ear-proximal scalp electrodes in a public dataset are not the same as true in-ear or cEEGrid recordings. Any finding transfers to real ear hardware only as an upper bound.
- L02Secondary analysis of existing datasets inherits their tasks, populations and recording conditions.
- L03Separation methods do not cleanly partition physiological sources; residual mixing is expected and will be reported rather than assumed away.
- L04Subject-independent performance is the number that matters for a real device and is usually much worse than subject-dependent performance.
- 01A measured degradation curve gives the Build section a defensible target for how few contacts an ear device can carry.
- 02If separated EMG contributes, the hardware requirement changes: contacts have to be placed for muscle as well as cortex.
- 03Whatever curve emerges becomes a realistic attribute set for the adoption study rather than an invented trade-off.
- 2026-09
- Research question fixed. Dataset selection in progress.